# AMD Ryzen 9 7950X3D (Zen 4), Debian 13, 4 CPUs (taskset -c 0-3), Python 3.12.15, scikit-learn 1.9.1, numpy 2.5.4, pandas 3.0.6, joblib 1.6.0, 12 October 2026
# lightgbm-takt 4.7.0 (free), library identical to the PyPI wheel (lib_lightgbm.so sha256 07b0da14...)

$ python pi_cost.py all work 3 ...
N = 4 CPUs, 3 rounds per setting, median wall time of the permutation_importance call
numpy, n_jobs=1, 1 thread                           18.04 s   5f242b6b5bd26ee0 same result
numpy, n_jobs=N, 1 thread                           10.17 s   5f242b6b5bd26ee0 same result
pandas, n_jobs=1, LightGBM default threads          10.95 s   5f242b6b5bd26ee0 same result
numpy, n_jobs=1, N threads                          10.93 s   5f242b6b5bd26ee0 same result

# alternating runs, official wheel (off) and lightgbm-takt (free): seconds, result hash, largest mean importance
# setting "numpy, n_jobs=1, N threads"
off 12.259 5f242b6b5bd26ee0 0.025123299949183765
free 11.155 5f242b6b5bd26ee0 0.025123299949183765
off 12.726 5f242b6b5bd26ee0 0.025123299949183765
free 10.908 5f242b6b5bd26ee0 0.025123299949183765
off 12.411 5f242b6b5bd26ee0 0.025123299949183765
free 11.094 5f242b6b5bd26ee0 0.025123299949183765
off 12.310 5f242b6b5bd26ee0 0.025123299949183765
free 10.874 5f242b6b5bd26ee0 0.025123299949183765
off 12.414 5f242b6b5bd26ee0 0.025123299949183765
free 10.961 5f242b6b5bd26ee0 0.025123299949183765